AptaBERT: Predicting aptamer binding interactions
Morsch, F.; Umasankar, I. L.; Sanz Moreta, L.; Latawa, P.; Lange, D. B.; Wengel, J.; Konjen, H.; Code, C.
Show abstract
AO_SCPLOWBSTRACTC_SCPLOWAptamers, short single-stranded DNA or RNA, are promising as future diagnostic and therapeutic agents. Traditional selection methods, such as the Systemic Evolution of Ligands by Exponential Enrichment (SELEX), are not without limitations being both resource-intensive and prone to biases in library construction and the selection phase. Leveraging Dianoxs extensive aptamer database, we introduce a novel computational approach, AptaBERT, built upon the BERT architecture. This method utilizes self-supervised pre-training on vast amounts of data, followed by supervised fine-tuning to enhance the prediction of aptamer interactions with proteins and small molecules. AptaBERT is fine-tuned for binary classification tasks, distinguishing between positive and negative interactions with proteins and small molecules. AptaBERT achieves a ROC-AUC of 96% for protein interactions, surpassing existing models by at least 15%. For small molecule interactions, AptaBERT attains an ROC-AUC of 85%. Our findings demonstrate AptaBERTs superior predictive capability and its potential to identify novel aptamers binding to targets.
Matching journals
The top 7 journals account for 50% of the predicted probability mass.
Similar papers in this journal
Similar papers in this journal
Similar papers in this journal
- Evaluating generalizability of artificial intelligence models for molecular datasets 95%
- Accelerating protein engineering with fitness landscape modeling and reinforcement learning 94%
- Improving protein function prediction with synthetic feature samples created by generative adversarial networks 94%
Similar papers in this journal
- GraphDTA: Predicting drug-target binding affinity with graph neural networks 95%
- Pair-EGRET: enhancing the prediction of protein-proteininteraction sites through graph attention networks and protein language models 95%
- Attention-based approach to predict drug-target interactions across seven target superfamilies 95%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.